Claude Opus 5.5 vs GPT-6 Astra

Claude Opus 5.5 by Anthropic · GPT-6 Astra by OpenAI · 28 shared benchmarks

Claude Opus 5.5

78.4

#2 · 100%

GPT-6 Astra

77.6

#3 · 100%

SI Score, rank and confidence side by side

Pillars — Claude Opus 5.5

Coding 72.5 / 64.8

Math 91.5 / 92.9

Preference 82.6 / 77.1

Reasoning 91.4 / 87.1

Right-hand value is GPT-6 Astra.

The basics

Attribute	Claude Opus 5.5	GPT-6 Astra
Input price / 1M	$4.00	$10.00
Output price / 1M	$20.00	$50.00
Context window	1M	1.1M
Released	Sep 22, 2026	Sep 3, 2026
Open weights	Closed	Closed

Shared benchmarks

28 in common

Only benchmarks both models have are compared. Raw values are in each benchmark's own unit; a direct comparison requires matching evaluation conditions. All reported source/variant rows are shown; no raw-result winner is assigned across unmatched harnesses. Sources and dates sit behind every dotted number.

Which should you choose?

What the data says — heuristics from the numbers above, not a verdict:

For reasoning, Claude Opus 5.5 leads by 4.3 normalized points.
For coding, Claude Opus 5.5 leads by 7.7 normalized points.
For preference, Claude Opus 5.5 leads by 5.6 normalized points.
Claude Opus 5.5 has the lower input-token price ($4.00 vs $10.00 per 1M).
Confidence differs: Claude Opus 5.5 100% vs GPT-6 Astra 100% — pending sources can still move either score.

Benchmark	Claude Opus 5.5	GPT-6 Astra	Conditions
arc agi v1 public eval	98.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-high 92.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-low 98.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] anthropic-claude-opus-5-5-max 98.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-medium 98.6%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-xhigh	98.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort. [variant] openai-gpt-6-astra-high 98.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort. [variant] openai-gpt-6-astra-low 97.8%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort. [variant] openai-gpt-6-astra-max 98.8%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort. [variant] openai-gpt-6-astra-medium 99%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort. [variant] openai-gpt-6-astra-xhigh	Check variant and harness
arc agi v1 semi private	98.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-high 88.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-low 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] anthropic-claude-opus-5-5-max 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-medium 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-xhigh	98.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort. [variant] openai-gpt-6-astra-high 96.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort. [variant] openai-gpt-6-astra-low 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort. [variant] openai-gpt-6-astra-max 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort. [variant] openai-gpt-6-astra-medium 98.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort. [variant] openai-gpt-6-astra-xhigh	Check variant and harness
arc agi v2 public eval	93.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-high 67.8%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-low 97.9%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] anthropic-claude-opus-5-5-max 93.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-medium 97.6%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-xhigh	97.9%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort. [variant] openai-gpt-6-astra-high 94.2%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort. [variant] openai-gpt-6-astra-low 97.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort. [variant] openai-gpt-6-astra-max 96.7%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort. [variant] openai-gpt-6-astra-medium 97.9%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort. [variant] openai-gpt-6-astra-xhigh	Check variant and harness
arc agi v2 semi private	93.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-high 70.1%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-low 91.7%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] anthropic-claude-opus-5-5-max 87.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-medium 92.5%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort using the standard harness. [variant] anthropic-claude-opus-5-5-xhigh	92.1%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort. [variant] openai-gpt-6-astra-high 85.4%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort. [variant] openai-gpt-6-astra-low 95%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort. [variant] openai-gpt-6-astra-max 92.1%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'medium' reasoning effort. [variant] openai-gpt-6-astra-medium 93.3%ARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'xhigh' reasoning effort. [variant] openai-gpt-6-astra-xhigh	Check variant and harness
frontiermath tier 4 v2	95%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	97.6%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] high 87.8%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] low 97.6%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max 97.6%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] medium 82.9%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] none 97.6%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] xhigh	Check variant and harness
frontiermath tiers 1 3 v2	91.2%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	93.7%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	Check variant and harness
gpqa diamond	90.6%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	95.8%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max 96%Lab-reported; metric accuracy; transcribed by MIT models.dev catalog; not independently evaluated [variant]	Check variant and harness
hle tools	67.7%Lab-reported; metric score; transcribed by MIT models.dev catalog; not independently evaluated [variant] max effort; with tools; production safeguards with fallback	57.2%Lab-reported; metric accuracy; transcribed by MIT models.dev catalog; not independently evaluated [variant] with tools	Check variant and harness
livebench coding code completion 2026_06_25	87.0%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	80.4%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench coding code generation 2026_06_25	91.5%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	80.3%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench coding javascript 2026_06_25	72.7%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	63.6%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench coding python 2026_06_25	70%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	65%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench coding typescript 2026_06_25	53.3%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	43.3%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench math amps hard 2026_06_25	99%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	98%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench math integrals with game 2026_06_25	99%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	100%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench math math comp 2026_06_25	97.1%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	97.1%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench math olympiad 2026_06_25	92.2%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	92.2%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench math simplify 2026_06_25	63%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	70.5%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning connections 2026_06_25	99.3%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	100%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning consecutive events 2026_06_25	90.4%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	90.6%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning logic with navigation 2026_06_25	80%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	88%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning spatial 2026_06_25	98%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	98%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning theory of mind 2026_06_25	84.6%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	84.6%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
livebench reasoning zebra puzzle 2026_06_25	100%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	100%LiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25	Check variant and harness
lmarena text	1511.7 eloPublished source fact [variant] text / overall	1442.3 eloPublished source fact [variant] text / overall	Check variant and harness
otis mock aime 2024 2025	100%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	100%Epoch-owned evaluation mean score; scores only, no benchmark questions [variant] max	Check variant and harness
terminal bench v0 1	58.7%Lab-reported; metric score; transcribed by MIT models.dev catalog; not independently evaluated [variant] max effort; production safeguards with fallback; 0.1	64.6%Lab-reported; metric success rate; transcribed by MIT models.dev catalog; not independently evaluated [variant] 0.1	Check variant and harness
terminal bench v4 0	66.4%Lab-reported; metric score; transcribed by MIT models.dev catalog; not independently evaluated [variant] xhigh effort; production safeguards with fallback; 4.0 64.8%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Claude Code; max	57.9%Lab-reported; metric success rate; transcribed by MIT models.dev catalog; not independently evaluated [variant] 4.0 57.9%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Codex; high 50.6%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Codex; low 58.2%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Codex; max 54.2%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Codex; medium 57.9%Terminal-Bench 4.0; published harness submission, 95% CI retained at source [variant] Codex; xhigh	Check variant and harness